app.py

import gradio as gr
from transformers import pipeline
import torch
from diffusers import StableDiffusionPipeline
import soundfile as sf
import speech_recognition as sr
import numpy as np
import os

# 初始化组件
# 使用较小的开源LLM进行提示增强
llm_pipe = pipeline("text-generation", model="mistralai/Mistral-7B-Instruct-v0.1")

# 初始化Stable Diffusion
sd_pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16
).to("cuda" if torch.cuda.is_available() else "cpu")

# 语音识别初始化
recognizer = sr.Recognizer()

def enhance_prompt(basic_prompt, style, detail_level, artist_style):
    """使用LLM增强提示词"""
    prompt_template = f"""
    根据以下简短描述创建一个详细的Stable Diffusion提示:
    原始描述: {basic_prompt}
    风格: {style}
    细节级别: {detail_level}
    艺术家风格: {artist_style}
    
    请生成一个包含以下元素的详细提示:
    - 主体描述
    - 环境/背景
    - 光照条件
    - 色彩风格
    - 艺术媒介(如数字绘画、油画等)
    - 质量描述(如4K、超详细等)
    
    生成的提示:
    """
    
    enhanced_prompt = llm_pipe(
        prompt_template,
        max_length=200,
        num_return_sequences=1,
        temperature=0.7
    )[0]['generated_text']
    
    # 清理生成的文本
    enhanced_prompt = enhanced_prompt.replace(prompt_template, "").strip()
    return enhanced_prompt

def generate_image(enhanced_prompt, steps, guidance_scale, seed):
    """使用Stable Diffusion生成图像"""
    if seed == -1:
        seed = torch.randint(0, 2**32, (1,)).item()
    
    generator = torch.Generator(device="cuda" if torch.cuda.is_available() else "cpu").manual_seed(seed)
    
    image = sd_pipe(
        enhanced_prompt,
        num_inference_steps=steps,
        guidance_scale=guidance_scale,
        generator=generator
    ).images[0]
    
    return image, seed

def process_audio(audio):
    """处理语音输入"""
    sr, audio_data = audio
    audio_array = np.array(audio_data, dtype=np.float32)
    
    # 保存临时文件供语音识别使用
    temp_file = "temp_audio.wav"
    sf.write(temp_file, audio_array, sr)
    
    with sr.AudioFile(temp_file) as source:
        audio_data = recognizer.record(source)
        try:
            text = recognizer.recognize_google(audio_data, language='en-US')
            os.remove(temp_file)
            return text
        except Exception as e:
            os.remove(temp_file)
            return f"语音识别错误: {str(e)}"

def full_process(basic_prompt, style, detail_level, artist_style, steps, guidance_scale, seed, use_audio, audio_input):
    """完整处理流程"""
    # 处理语音输入
    if use_audio and audio_input is not None:
        basic_prompt = process_audio(audio_input)
    
    # 生成增强提示
    enhanced_prompt = enhance_prompt(basic_prompt, style, detail_level, artist_style)
    
    # 生成图像
    image, used_seed = generate_image(enhanced_prompt, steps, guidance_scale, seed)
    
    return enhanced_prompt, image, used_seed

# Gradio界面
with gr.Blocks(title="魔法树屋图像生成器") as demo:
    gr.Markdown("# 🎨 魔法树屋图像生成器")
    gr.Markdown("输入简短描述或使用语音输入，生成精美图像！")
    
    with gr.Row():
        with gr.Column():
            # 输入部分
            use_audio = gr.Checkbox(label="使用语音输入")
            audio_input = gr.Audio(label="录音", visible=False)
            
            basic_prompt = gr.Textbox(
                label="简短描述",
                placeholder="例如: 天空中的魔法树屋",
                visible=True
            )
            
            # 当复选框变化时切换输入方式
            def toggle_input(use_audio):
                return {
                    basic_prompt: gr.update(visible=not use_audio),
                    audio_input: gr.update(visible=use_audio)
                }
            
            use_audio.change(
                toggle_input,
                inputs=use_audio,
                outputs=[basic_prompt, audio_input]
            )
            
            # 风格选项
            style = gr.Dropdown(
                label="风格",
                choices=["现实主义", "幻想艺术", "赛博朋克", "水墨画", "卡通", "极简主义"],
                value="幻想艺术"
            )
            
            detail_level = gr.Slider(
                label="细节级别",
                minimum=1,
                maximum=5,
                step=1,
                value=3
            )
            
            artist_style = gr.Dropdown(
                label="艺术家风格",
                choices=["无", "梵高", "毕加索", "莫奈", "达利", "宫崎骏"],
                value="无"
            )
            
            # 高级选项
            with gr.Accordion("高级选项", open=False):
                steps = gr.Slider(
                    label="生成步数",
                    minimum=20,
                    maximum=100,
                    step=5,
                    value=50
                )
                
                guidance_scale = gr.Slider(
                    label="引导尺度",
                    minimum=1.0,
                    maximum=20.0,
                    step=0.5,
                    value=7.5
                )
                
                seed = gr.Number(
                    label="随机种子 (-1 表示随机)",
                    value=-1
                )
            
            submit_btn = gr.Button("生成图像", variant="primary")
        
        with gr.Column():
            # 输出部分
            enhanced_prompt = gr.Textbox(
                label="生成的提示",
                interactive=False
            )
            
            image_output = gr.Image(
                label="生成的图像",
                height=512
            )
            
            used_seed = gr.Number(
                label="使用的种子",
                interactive=False
            )
    
    # 连接按钮
    submit_btn.click(
        fn=full_process,
        inputs=[
            basic_prompt, style, detail_level, artist_style,
            steps, guidance_scale, seed, use_audio, audio_input
        ],
        outputs=[enhanced_prompt, image_output, used_seed]
    )

# 对于Hugging Face Spaces，我们需要设置队列
demo.queue()

# 启动应用
if __name__ == "__main__":
    demo.launch()